Skip to content

fix(vllm): warm Inkling convolution caches with native prefills #15257

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Arsene12358 wants to merge 16 commits into
base: main
Choose a base branch
from
Open

fix(vllm): warm Inkling convolution caches with native prefills #15257

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
42a3a26
fix(vllm): warm finite-window benchmark caches with native prefills
Arsene12358 Sep 24, 2026
59bdde0
fix(vllm): declare the warm-up plan type for mypy
Arsene12358 Oct 1, 2026
609c223
fix(vllm): keep native plan numbering valid when attention-DP drops p…
Arsene12358 Oct 1, 2026
b5d25b8
fix(vllm): keep dense attention-only models on synthetic KV
Arsene12358 Oct 2, 2026
1338e23
fix(vllm): warm Mamba and KDA state layouts with native prefills
Arsene12358 Oct 2, 2026
5a0456e
fix(vllm): reserve align headroom for native stages, fix live-state help
Arsene12358 Oct 2, 2026
1f82903
fix(vllm): label gate-skipped decode points skip:<reason>
Arsene12358 Oct 2, 2026
0a6544e
test(vllm): tidy native warm-up tests and document the native cost
Arsene12358 Oct 2, 2026
4d72a2a
fix(vllm): address the final review of the native warm-up update
Arsene12358 Oct 2, 2026
211233c
Merge branch 'main' into fix/inkling-real-kv-warmup
Arsene12358 Oct 3, 2026
60da85a
Merge branch 'main' into fix/inkling-real-kv-warmup
Arsene12358 Oct 10, 2026
1fa6362
test(vllm): give the native grid-numbering stub the state #15110 reads
Arsene12358 Oct 10, 2026
af3c119
fix(vllm): record prompt evidence for native warm-up decode points
Arsene12358 Oct 10, 2026
e55735b
test(vllm): run the native admission-cap test with vLLM 0.31.0's name
Arsene12358 Oct 10, 2026
8d484b4
docs(vllm): describe native warm-up in the benchmark evidence docs
Arsene12358 Oct 10, 2026
5d654ec
docs(vllm): limit the native prompt-length note to real_kv rows
Arsene12358 Oct 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions components/src/dynamo/vllm/backend_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -491,8 +491,9 @@ def add_arguments(self, parser) -> None:
"recurrent-state groups forked from the parked chain's live state block "
"instead of skipping the warm-up. Attention KV is the chain's real prefix; "
"the recurrent state is a valid but deeper-context state (shallow points "
"read a few percent fast). Mutually exclusive with "
"--benchmark-randomize-kda-state."
"read a few percent fast). Applies only to layouts that do not qualify "
"for native exact-context warm-up, which takes precedence. Mutually "
"exclusive with --benchmark-randomize-kda-state."
),
)
add_argument(
Expand Down
Loading
Loading